1 /* 2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "gc/shared/barrierSet.hpp" 27 #include "gc/shared/c2/barrierSetC2.hpp" 28 #include "memory/allocation.inline.hpp" 29 #include "memory/resourceArea.hpp" 30 #include "opto/block.hpp" 31 #include "opto/callnode.hpp" 32 #include "opto/castnode.hpp" 33 #include "opto/cfgnode.hpp" 34 #include "opto/idealGraphPrinter.hpp" 35 #include "opto/loopnode.hpp" 36 #include "opto/machnode.hpp" 37 #include "opto/opcodes.hpp" 38 #include "opto/phaseX.hpp" 39 #include "opto/regalloc.hpp" 40 #include "opto/rootnode.hpp" 41 #include "utilities/macros.hpp" 42 #include "utilities/powerOfTwo.hpp" 43 44 //============================================================================= 45 #define NODE_HASH_MINIMUM_SIZE 255 46 47 //------------------------------NodeHash--------------------------------------- 48 NodeHash::NodeHash(Arena *arena, uint est_max_size) : 49 _a(arena), 50 _max( round_up(est_max_size < NODE_HASH_MINIMUM_SIZE ? NODE_HASH_MINIMUM_SIZE : est_max_size) ), 51 _inserts(0), _insert_limit( insert_limit() ), 52 _table( NEW_ARENA_ARRAY( _a , Node* , _max ) ) 53 #ifndef PRODUCT 54 , _grows(0),_look_probes(0), _lookup_hits(0), _lookup_misses(0), 55 _insert_probes(0), _delete_probes(0), _delete_hits(0), _delete_misses(0), 56 _total_inserts(0), _total_insert_probes(0) 57 #endif 58 { 59 // _sentinel must be in the current node space 60 _sentinel = new ProjNode(nullptr, TypeFunc::Control); 61 memset(_table,0,sizeof(Node*)*_max); 62 } 63 64 //------------------------------hash_find-------------------------------------- 65 // Find in hash table 66 Node *NodeHash::hash_find( const Node *n ) { 67 // ((Node*)n)->set_hash( n->hash() ); 68 uint hash = n->hash(); 69 if (hash == Node::NO_HASH) { 70 NOT_PRODUCT( _lookup_misses++ ); 71 return nullptr; 72 } 73 uint key = hash & (_max-1); 74 uint stride = key | 0x01; 75 NOT_PRODUCT( _look_probes++ ); 76 Node *k = _table[key]; // Get hashed value 77 if( !k ) { // ?Miss? 78 NOT_PRODUCT( _lookup_misses++ ); 79 return nullptr; // Miss! 80 } 81 82 int op = n->Opcode(); 83 uint req = n->req(); 84 while( 1 ) { // While probing hash table 85 if( k->req() == req && // Same count of inputs 86 k->Opcode() == op ) { // Same Opcode 87 for( uint i=0; i<req; i++ ) 88 if( n->in(i)!=k->in(i)) // Different inputs? 89 goto collision; // "goto" is a speed hack... 90 if( n->cmp(*k) ) { // Check for any special bits 91 NOT_PRODUCT( _lookup_hits++ ); 92 return k; // Hit! 93 } 94 } 95 collision: 96 NOT_PRODUCT( _look_probes++ ); 97 key = (key + stride/*7*/) & (_max-1); // Stride through table with relative prime 98 k = _table[key]; // Get hashed value 99 if( !k ) { // ?Miss? 100 NOT_PRODUCT( _lookup_misses++ ); 101 return nullptr; // Miss! 102 } 103 } 104 ShouldNotReachHere(); 105 return nullptr; 106 } 107 108 //------------------------------hash_find_insert------------------------------- 109 // Find in hash table, insert if not already present 110 // Used to preserve unique entries in hash table 111 Node *NodeHash::hash_find_insert( Node *n ) { 112 // n->set_hash( ); 113 uint hash = n->hash(); 114 if (hash == Node::NO_HASH) { 115 NOT_PRODUCT( _lookup_misses++ ); 116 return nullptr; 117 } 118 uint key = hash & (_max-1); 119 uint stride = key | 0x01; // stride must be relatively prime to table siz 120 uint first_sentinel = 0; // replace a sentinel if seen. 121 NOT_PRODUCT( _look_probes++ ); 122 Node *k = _table[key]; // Get hashed value 123 if( !k ) { // ?Miss? 124 NOT_PRODUCT( _lookup_misses++ ); 125 _table[key] = n; // Insert into table! 126 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 127 check_grow(); // Grow table if insert hit limit 128 return nullptr; // Miss! 129 } 130 else if( k == _sentinel ) { 131 first_sentinel = key; // Can insert here 132 } 133 134 int op = n->Opcode(); 135 uint req = n->req(); 136 while( 1 ) { // While probing hash table 137 if( k->req() == req && // Same count of inputs 138 k->Opcode() == op ) { // Same Opcode 139 for( uint i=0; i<req; i++ ) 140 if( n->in(i)!=k->in(i)) // Different inputs? 141 goto collision; // "goto" is a speed hack... 142 if( n->cmp(*k) ) { // Check for any special bits 143 NOT_PRODUCT( _lookup_hits++ ); 144 return k; // Hit! 145 } 146 } 147 collision: 148 NOT_PRODUCT( _look_probes++ ); 149 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 150 k = _table[key]; // Get hashed value 151 if( !k ) { // ?Miss? 152 NOT_PRODUCT( _lookup_misses++ ); 153 key = (first_sentinel == 0) ? key : first_sentinel; // ?saw sentinel? 154 _table[key] = n; // Insert into table! 155 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 156 check_grow(); // Grow table if insert hit limit 157 return nullptr; // Miss! 158 } 159 else if( first_sentinel == 0 && k == _sentinel ) { 160 first_sentinel = key; // Can insert here 161 } 162 163 } 164 ShouldNotReachHere(); 165 return nullptr; 166 } 167 168 //------------------------------hash_insert------------------------------------ 169 // Insert into hash table 170 void NodeHash::hash_insert( Node *n ) { 171 // // "conflict" comments -- print nodes that conflict 172 // bool conflict = false; 173 // n->set_hash(); 174 uint hash = n->hash(); 175 if (hash == Node::NO_HASH) { 176 return; 177 } 178 check_grow(); 179 uint key = hash & (_max-1); 180 uint stride = key | 0x01; 181 182 while( 1 ) { // While probing hash table 183 NOT_PRODUCT( _insert_probes++ ); 184 Node *k = _table[key]; // Get hashed value 185 if( !k || (k == _sentinel) ) break; // Found a slot 186 assert( k != n, "already inserted" ); 187 // if( PrintCompilation && PrintOptoStatistics && Verbose ) { tty->print(" conflict: "); k->dump(); conflict = true; } 188 key = (key + stride) & (_max-1); // Stride through table w/ relative prime 189 } 190 _table[key] = n; // Insert into table! 191 debug_only(n->enter_hash_lock()); // Lock down the node while in the table. 192 // if( conflict ) { n->dump(); } 193 } 194 195 //------------------------------hash_delete------------------------------------ 196 // Replace in hash table with sentinel 197 bool NodeHash::hash_delete( const Node *n ) { 198 Node *k; 199 uint hash = n->hash(); 200 if (hash == Node::NO_HASH) { 201 NOT_PRODUCT( _delete_misses++ ); 202 return false; 203 } 204 uint key = hash & (_max-1); 205 uint stride = key | 0x01; 206 debug_only( uint counter = 0; ); 207 for( ; /* (k != nullptr) && (k != _sentinel) */; ) { 208 debug_only( counter++ ); 209 NOT_PRODUCT( _delete_probes++ ); 210 k = _table[key]; // Get hashed value 211 if( !k ) { // Miss? 212 NOT_PRODUCT( _delete_misses++ ); 213 return false; // Miss! Not in chain 214 } 215 else if( n == k ) { 216 NOT_PRODUCT( _delete_hits++ ); 217 _table[key] = _sentinel; // Hit! Label as deleted entry 218 debug_only(((Node*)n)->exit_hash_lock()); // Unlock the node upon removal from table. 219 return true; 220 } 221 else { 222 // collision: move through table with prime offset 223 key = (key + stride/*7*/) & (_max-1); 224 assert( counter <= _insert_limit, "Cycle in hash-table"); 225 } 226 } 227 ShouldNotReachHere(); 228 return false; 229 } 230 231 //------------------------------round_up--------------------------------------- 232 // Round up to nearest power of 2 233 uint NodeHash::round_up(uint x) { 234 x += (x >> 2); // Add 25% slop 235 return MAX2(16U, round_up_power_of_2(x)); 236 } 237 238 //------------------------------grow------------------------------------------- 239 // Grow _table to next power of 2 and insert old entries 240 void NodeHash::grow() { 241 // Record old state 242 uint old_max = _max; 243 Node **old_table = _table; 244 // Construct new table with twice the space 245 #ifndef PRODUCT 246 _grows++; 247 _total_inserts += _inserts; 248 _total_insert_probes += _insert_probes; 249 _insert_probes = 0; 250 #endif 251 _inserts = 0; 252 _max = _max << 1; 253 _table = NEW_ARENA_ARRAY( _a , Node* , _max ); // (Node**)_a->Amalloc( _max * sizeof(Node*) ); 254 memset(_table,0,sizeof(Node*)*_max); 255 _insert_limit = insert_limit(); 256 // Insert old entries into the new table 257 for( uint i = 0; i < old_max; i++ ) { 258 Node *m = *old_table++; 259 if( !m || m == _sentinel ) continue; 260 debug_only(m->exit_hash_lock()); // Unlock the node upon removal from old table. 261 hash_insert(m); 262 } 263 } 264 265 //------------------------------clear------------------------------------------ 266 // Clear all entries in _table to null but keep storage 267 void NodeHash::clear() { 268 #ifdef ASSERT 269 // Unlock all nodes upon removal from table. 270 for (uint i = 0; i < _max; i++) { 271 Node* n = _table[i]; 272 if (!n || n == _sentinel) continue; 273 n->exit_hash_lock(); 274 } 275 #endif 276 277 memset( _table, 0, _max * sizeof(Node*) ); 278 } 279 280 //-----------------------remove_useless_nodes---------------------------------- 281 // Remove useless nodes from value table, 282 // implementation does not depend on hash function 283 void NodeHash::remove_useless_nodes(VectorSet &useful) { 284 285 // Dead nodes in the hash table inherited from GVN should not replace 286 // existing nodes, remove dead nodes. 287 uint max = size(); 288 Node *sentinel_node = sentinel(); 289 for( uint i = 0; i < max; ++i ) { 290 Node *n = at(i); 291 if(n != nullptr && n != sentinel_node && !useful.test(n->_idx)) { 292 debug_only(n->exit_hash_lock()); // Unlock the node when removed 293 _table[i] = sentinel_node; // Replace with placeholder 294 } 295 } 296 } 297 298 299 void NodeHash::check_no_speculative_types() { 300 #ifdef ASSERT 301 uint max = size(); 302 Unique_Node_List live_nodes; 303 Compile::current()->identify_useful_nodes(live_nodes); 304 Node *sentinel_node = sentinel(); 305 for (uint i = 0; i < max; ++i) { 306 Node *n = at(i); 307 if (n != nullptr && 308 n != sentinel_node && 309 n->is_Type() && 310 live_nodes.member(n)) { 311 TypeNode* tn = n->as_Type(); 312 const Type* t = tn->type(); 313 const Type* t_no_spec = t->remove_speculative(); 314 assert(t == t_no_spec, "dead node in hash table or missed node during speculative cleanup"); 315 } 316 } 317 #endif 318 } 319 320 #ifndef PRODUCT 321 //------------------------------dump------------------------------------------- 322 // Dump statistics for the hash table 323 void NodeHash::dump() { 324 _total_inserts += _inserts; 325 _total_insert_probes += _insert_probes; 326 if (PrintCompilation && PrintOptoStatistics && Verbose && (_inserts > 0)) { 327 if (WizardMode) { 328 for (uint i=0; i<_max; i++) { 329 if (_table[i]) 330 tty->print("%d/%d/%d ",i,_table[i]->hash()&(_max-1),_table[i]->_idx); 331 } 332 } 333 tty->print("\nGVN Hash stats: %d grows to %d max_size\n", _grows, _max); 334 tty->print(" %d/%d (%8.1f%% full)\n", _inserts, _max, (double)_inserts/_max*100.0); 335 tty->print(" %dp/(%dh+%dm) (%8.2f probes/lookup)\n", _look_probes, _lookup_hits, _lookup_misses, (double)_look_probes/(_lookup_hits+_lookup_misses)); 336 tty->print(" %dp/%di (%8.2f probes/insert)\n", _total_insert_probes, _total_inserts, (double)_total_insert_probes/_total_inserts); 337 // sentinels increase lookup cost, but not insert cost 338 assert((_lookup_misses+_lookup_hits)*4+100 >= _look_probes, "bad hash function"); 339 assert( _inserts+(_inserts>>3) < _max, "table too full" ); 340 assert( _inserts*3+100 >= _insert_probes, "bad hash function" ); 341 } 342 } 343 344 Node *NodeHash::find_index(uint idx) { // For debugging 345 // Find an entry by its index value 346 for( uint i = 0; i < _max; i++ ) { 347 Node *m = _table[i]; 348 if( !m || m == _sentinel ) continue; 349 if( m->_idx == (uint)idx ) return m; 350 } 351 return nullptr; 352 } 353 #endif 354 355 #ifdef ASSERT 356 NodeHash::~NodeHash() { 357 // Unlock all nodes upon destruction of table. 358 if (_table != (Node**)badAddress) clear(); 359 } 360 #endif 361 362 363 //============================================================================= 364 //------------------------------PhaseRemoveUseless----------------------------- 365 // 1) Use a breadthfirst walk to collect useful nodes reachable from root. 366 PhaseRemoveUseless::PhaseRemoveUseless(PhaseGVN* gvn, Unique_Node_List& worklist, PhaseNumber phase_num) : Phase(phase_num) { 367 C->print_method(PHASE_BEFORE_REMOVEUSELESS, 3); 368 // Implementation requires an edge from root to each SafePointNode 369 // at a backward branch. Inserted in add_safepoint(). 370 371 // Identify nodes that are reachable from below, useful. 372 C->identify_useful_nodes(_useful); 373 // Update dead node list 374 C->update_dead_node_list(_useful); 375 376 // Remove all useless nodes from PhaseValues' recorded types 377 // Must be done before disconnecting nodes to preserve hash-table-invariant 378 gvn->remove_useless_nodes(_useful.member_set()); 379 380 // Remove all useless nodes from future worklist 381 worklist.remove_useless_nodes(_useful.member_set()); 382 383 // Disconnect 'useless' nodes that are adjacent to useful nodes 384 C->disconnect_useless_nodes(_useful, worklist); 385 } 386 387 //============================================================================= 388 //------------------------------PhaseRenumberLive------------------------------ 389 // First, remove useless nodes (equivalent to identifying live nodes). 390 // Then, renumber live nodes. 391 // 392 // The set of live nodes is returned by PhaseRemoveUseless in the _useful structure. 393 // If the number of live nodes is 'x' (where 'x' == _useful.size()), then the 394 // PhaseRenumberLive updates the node ID of each node (the _idx field) with a unique 395 // value in the range [0, x). 396 // 397 // At the end of the PhaseRenumberLive phase, the compiler's count of unique nodes is 398 // updated to 'x' and the list of dead nodes is reset (as there are no dead nodes). 399 // 400 // The PhaseRenumberLive phase updates two data structures with the new node IDs. 401 // (1) The "worklist" is "C->igvn_worklist()", which is to collect which nodes need to 402 // be processed by IGVN after removal of the useless nodes. 403 // (2) Type information "gvn->types()" (same as "C->types()") maps every node ID to 404 // the node's type. The mapping is updated to use the new node IDs as well. We 405 // create a new map, and swap it with the old one. 406 // 407 // Other data structures used by the compiler are not updated. The hash table for value 408 // numbering ("C->node_hash()", referenced by PhaseValue::_table) is not updated because 409 // computing the hash values is not based on node IDs. 410 PhaseRenumberLive::PhaseRenumberLive(PhaseGVN* gvn, 411 Unique_Node_List& worklist, 412 PhaseNumber phase_num) : 413 PhaseRemoveUseless(gvn, worklist, Remove_Useless_And_Renumber_Live), 414 _new_type_array(C->comp_arena()), 415 _old2new_map(C->unique(), C->unique(), -1), 416 _is_pass_finished(false), 417 _live_node_count(C->live_nodes()) 418 { 419 assert(RenumberLiveNodes, "RenumberLiveNodes must be set to true for node renumbering to take place"); 420 assert(C->live_nodes() == _useful.size(), "the number of live nodes must match the number of useful nodes"); 421 assert(_delayed.size() == 0, "should be empty"); 422 assert(&worklist == C->igvn_worklist(), "reference still same as the one from Compile"); 423 assert(&gvn->types() == C->types(), "reference still same as that from Compile"); 424 425 GrowableArray<Node_Notes*>* old_node_note_array = C->node_note_array(); 426 if (old_node_note_array != nullptr) { 427 int new_size = (_useful.size() >> 8) + 1; // The node note array uses blocks, see C->_log2_node_notes_block_size 428 new_size = MAX2(8, new_size); 429 C->set_node_note_array(new (C->comp_arena()) GrowableArray<Node_Notes*> (C->comp_arena(), new_size, 0, nullptr)); 430 C->grow_node_notes(C->node_note_array(), new_size); 431 } 432 433 assert(worklist.is_subset_of(_useful), "only useful nodes should still be in the worklist"); 434 435 // Iterate over the set of live nodes. 436 for (uint current_idx = 0; current_idx < _useful.size(); current_idx++) { 437 Node* n = _useful.at(current_idx); 438 439 const Type* type = gvn->type_or_null(n); 440 _new_type_array.map(current_idx, type); 441 442 assert(_old2new_map.at(n->_idx) == -1, "already seen"); 443 _old2new_map.at_put(n->_idx, current_idx); 444 445 if (old_node_note_array != nullptr) { 446 Node_Notes* nn = C->locate_node_notes(old_node_note_array, n->_idx); 447 C->set_node_notes_at(current_idx, nn); 448 } 449 450 n->set_idx(current_idx); // Update node ID. 451 452 if (update_embedded_ids(n) < 0) { 453 _delayed.push(n); // has embedded IDs; handle later 454 } 455 } 456 457 // VectorSet in Unique_Node_Set must be recomputed, since IDs have changed. 458 worklist.recompute_idx_set(); 459 460 assert(_live_node_count == _useful.size(), "all live nodes must be processed"); 461 462 _is_pass_finished = true; // pass finished; safe to process delayed updates 463 464 while (_delayed.size() > 0) { 465 Node* n = _delayed.pop(); 466 int no_of_updates = update_embedded_ids(n); 467 assert(no_of_updates > 0, "should be updated"); 468 } 469 470 // Replace the compiler's type information with the updated type information. 471 gvn->types().swap(_new_type_array); 472 473 // Update the unique node count of the compilation to the number of currently live nodes. 474 C->set_unique(_live_node_count); 475 476 // Set the dead node count to 0 and reset dead node list. 477 C->reset_dead_node_list(); 478 } 479 480 int PhaseRenumberLive::new_index(int old_idx) { 481 assert(_is_pass_finished, "not finished"); 482 if (_old2new_map.at(old_idx) == -1) { // absent 483 // Allocate a placeholder to preserve uniqueness 484 _old2new_map.at_put(old_idx, _live_node_count); 485 _live_node_count++; 486 } 487 return _old2new_map.at(old_idx); 488 } 489 490 int PhaseRenumberLive::update_embedded_ids(Node* n) { 491 int no_of_updates = 0; 492 if (n->is_Phi()) { 493 PhiNode* phi = n->as_Phi(); 494 if (phi->_inst_id != -1) { 495 if (!_is_pass_finished) { 496 return -1; // delay 497 } 498 int new_idx = new_index(phi->_inst_id); 499 assert(new_idx != -1, ""); 500 phi->_inst_id = new_idx; 501 no_of_updates++; 502 } 503 if (phi->_inst_mem_id != -1) { 504 if (!_is_pass_finished) { 505 return -1; // delay 506 } 507 int new_idx = new_index(phi->_inst_mem_id); 508 assert(new_idx != -1, ""); 509 phi->_inst_mem_id = new_idx; 510 no_of_updates++; 511 } 512 } 513 514 const Type* type = _new_type_array.fast_lookup(n->_idx); 515 if (type != nullptr && type->isa_oopptr() && type->is_oopptr()->is_known_instance()) { 516 if (!_is_pass_finished) { 517 return -1; // delay 518 } 519 int old_idx = type->is_oopptr()->instance_id(); 520 int new_idx = new_index(old_idx); 521 const Type* new_type = type->is_oopptr()->with_instance_id(new_idx); 522 _new_type_array.map(n->_idx, new_type); 523 no_of_updates++; 524 } 525 526 return no_of_updates; 527 } 528 529 void PhaseValues::init_con_caches() { 530 memset(_icons,0,sizeof(_icons)); 531 memset(_lcons,0,sizeof(_lcons)); 532 memset(_zcons,0,sizeof(_zcons)); 533 } 534 535 //--------------------------------find_int_type-------------------------------- 536 const TypeInt* PhaseValues::find_int_type(Node* n) { 537 if (n == nullptr) return nullptr; 538 // Call type_or_null(n) to determine node's type since we might be in 539 // parse phase and call n->Value() may return wrong type. 540 // (For example, a phi node at the beginning of loop parsing is not ready.) 541 const Type* t = type_or_null(n); 542 if (t == nullptr) return nullptr; 543 return t->isa_int(); 544 } 545 546 547 //-------------------------------find_long_type-------------------------------- 548 const TypeLong* PhaseValues::find_long_type(Node* n) { 549 if (n == nullptr) return nullptr; 550 // (See comment above on type_or_null.) 551 const Type* t = type_or_null(n); 552 if (t == nullptr) return nullptr; 553 return t->isa_long(); 554 } 555 556 //------------------------------~PhaseValues----------------------------------- 557 #ifndef PRODUCT 558 PhaseValues::~PhaseValues() { 559 // Statistics for NodeHash 560 _table.dump(); 561 // Statistics for value progress and efficiency 562 if( PrintCompilation && Verbose && WizardMode ) { 563 tty->print("\n%sValues: %d nodes ---> %d/%d (%d)", 564 is_IterGVN() ? "Iter" : " ", C->unique(), made_progress(), made_transforms(), made_new_values()); 565 if( made_transforms() != 0 ) { 566 tty->print_cr(" ratio %f", made_progress()/(float)made_transforms() ); 567 } else { 568 tty->cr(); 569 } 570 } 571 } 572 #endif 573 574 //------------------------------makecon---------------------------------------- 575 ConNode* PhaseValues::makecon(const Type* t) { 576 assert(t->singleton(), "must be a constant"); 577 assert(!t->empty() || t == Type::TOP, "must not be vacuous range"); 578 switch (t->base()) { // fast paths 579 case Type::Half: 580 case Type::Top: return (ConNode*) C->top(); 581 case Type::Int: return intcon( t->is_int()->get_con() ); 582 case Type::Long: return longcon( t->is_long()->get_con() ); 583 default: break; 584 } 585 if (t->is_zero_type()) 586 return zerocon(t->basic_type()); 587 return uncached_makecon(t); 588 } 589 590 //--------------------------uncached_makecon----------------------------------- 591 // Make an idealized constant - one of ConINode, ConPNode, etc. 592 ConNode* PhaseValues::uncached_makecon(const Type *t) { 593 assert(t->singleton(), "must be a constant"); 594 ConNode* x = ConNode::make(t); 595 ConNode* k = (ConNode*)hash_find_insert(x); // Value numbering 596 if (k == nullptr) { 597 set_type(x, t); // Missed, provide type mapping 598 GrowableArray<Node_Notes*>* nna = C->node_note_array(); 599 if (nna != nullptr) { 600 Node_Notes* loc = C->locate_node_notes(nna, x->_idx, true); 601 loc->clear(); // do not put debug info on constants 602 } 603 } else { 604 x->destruct(this); // Hit, destroy duplicate constant 605 x = k; // use existing constant 606 } 607 return x; 608 } 609 610 //------------------------------intcon----------------------------------------- 611 // Fast integer constant. Same as "transform(new ConINode(TypeInt::make(i)))" 612 ConINode* PhaseValues::intcon(jint i) { 613 // Small integer? Check cache! Check that cached node is not dead 614 if (i >= _icon_min && i <= _icon_max) { 615 ConINode* icon = _icons[i-_icon_min]; 616 if (icon != nullptr && icon->in(TypeFunc::Control) != nullptr) 617 return icon; 618 } 619 ConINode* icon = (ConINode*) uncached_makecon(TypeInt::make(i)); 620 assert(icon->is_Con(), ""); 621 if (i >= _icon_min && i <= _icon_max) 622 _icons[i-_icon_min] = icon; // Cache small integers 623 return icon; 624 } 625 626 //------------------------------longcon---------------------------------------- 627 // Fast long constant. 628 ConLNode* PhaseValues::longcon(jlong l) { 629 // Small integer? Check cache! Check that cached node is not dead 630 if (l >= _lcon_min && l <= _lcon_max) { 631 ConLNode* lcon = _lcons[l-_lcon_min]; 632 if (lcon != nullptr && lcon->in(TypeFunc::Control) != nullptr) 633 return lcon; 634 } 635 ConLNode* lcon = (ConLNode*) uncached_makecon(TypeLong::make(l)); 636 assert(lcon->is_Con(), ""); 637 if (l >= _lcon_min && l <= _lcon_max) 638 _lcons[l-_lcon_min] = lcon; // Cache small integers 639 return lcon; 640 } 641 ConNode* PhaseValues::integercon(jlong l, BasicType bt) { 642 if (bt == T_INT) { 643 return intcon(checked_cast<jint>(l)); 644 } 645 assert(bt == T_LONG, "not an integer"); 646 return longcon(l); 647 } 648 649 650 //------------------------------zerocon----------------------------------------- 651 // Fast zero or null constant. Same as "transform(ConNode::make(Type::get_zero_type(bt)))" 652 ConNode* PhaseValues::zerocon(BasicType bt) { 653 assert((uint)bt <= _zcon_max, "domain check"); 654 ConNode* zcon = _zcons[bt]; 655 if (zcon != nullptr && zcon->in(TypeFunc::Control) != nullptr) 656 return zcon; 657 zcon = (ConNode*) uncached_makecon(Type::get_zero_type(bt)); 658 _zcons[bt] = zcon; 659 return zcon; 660 } 661 662 663 664 //============================================================================= 665 Node* PhaseGVN::apply_ideal(Node* k, bool can_reshape) { 666 Node* i = BarrierSet::barrier_set()->barrier_set_c2()->ideal_node(this, k, can_reshape); 667 if (i == nullptr) { 668 i = k->Ideal(this, can_reshape); 669 } 670 return i; 671 } 672 673 //------------------------------transform-------------------------------------- 674 // Return a node which computes the same function as this node, but in a 675 // faster or cheaper fashion. 676 Node *PhaseGVN::transform( Node *n ) { 677 return transform_no_reclaim(n); 678 } 679 680 //------------------------------transform-------------------------------------- 681 // Return a node which computes the same function as this node, but 682 // in a faster or cheaper fashion. 683 Node *PhaseGVN::transform_no_reclaim(Node *n) { 684 NOT_PRODUCT( set_transforms(); ) 685 686 // Apply the Ideal call in a loop until it no longer applies 687 Node* k = n; 688 Node* i = apply_ideal(k, /*can_reshape=*/false); 689 NOT_PRODUCT(uint loop_count = 1;) 690 while (i != nullptr) { 691 assert(i->_idx >= k->_idx, "Idealize should return new nodes, use Identity to return old nodes" ); 692 k = i; 693 #ifdef ASSERT 694 if (loop_count >= K + C->live_nodes()) { 695 dump_infinite_loop_info(i, "PhaseGVN::transform_no_reclaim"); 696 } 697 #endif 698 i = apply_ideal(k, /*can_reshape=*/false); 699 NOT_PRODUCT(loop_count++;) 700 } 701 NOT_PRODUCT(if (loop_count != 0) { set_progress(); }) 702 703 // If brand new node, make space in type array. 704 ensure_type_or_null(k); 705 706 // Since I just called 'Value' to compute the set of run-time values 707 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 708 // cache Value. Later requests for the local phase->type of this Node can 709 // use the cached Value instead of suffering with 'bottom_type'. 710 const Type* t = k->Value(this); // Get runtime Value set 711 assert(t != nullptr, "value sanity"); 712 if (type_or_null(k) != t) { 713 #ifndef PRODUCT 714 // Do not count initial visit to node as a transformation 715 if (type_or_null(k) == nullptr) { 716 inc_new_values(); 717 set_progress(); 718 } 719 #endif 720 set_type(k, t); 721 // If k is a TypeNode, capture any more-precise type permanently into Node 722 k->raise_bottom_type(t); 723 } 724 725 if (t->singleton() && !k->is_Con()) { 726 NOT_PRODUCT(set_progress();) 727 return makecon(t); // Turn into a constant 728 } 729 730 // Now check for Identities 731 i = k->Identity(this); // Look for a nearby replacement 732 if (i != k) { // Found? Return replacement! 733 NOT_PRODUCT(set_progress();) 734 return i; 735 } 736 737 // Global Value Numbering 738 i = hash_find_insert(k); // Insert if new 739 if (i && (i != k)) { 740 // Return the pre-existing node 741 NOT_PRODUCT(set_progress();) 742 return i; 743 } 744 745 // Return Idealized original 746 return k; 747 } 748 749 bool PhaseGVN::is_dominator_helper(Node *d, Node *n, bool linear_only) { 750 if (d->is_top() || (d->is_Proj() && d->in(0)->is_top())) { 751 return false; 752 } 753 if (n->is_top() || (n->is_Proj() && n->in(0)->is_top())) { 754 return false; 755 } 756 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 757 int i = 0; 758 while (d != n) { 759 n = IfNode::up_one_dom(n, linear_only); 760 i++; 761 if (n == nullptr || i >= 100) { 762 return false; 763 } 764 } 765 return true; 766 } 767 768 #ifdef ASSERT 769 //------------------------------dead_loop_check-------------------------------- 770 // Check for a simple dead loop when a data node references itself directly 771 // or through an other data node excluding cons and phis. 772 void PhaseGVN::dead_loop_check( Node *n ) { 773 // Phi may reference itself in a loop 774 if (n != nullptr && !n->is_dead_loop_safe() && !n->is_CFG()) { 775 // Do 2 levels check and only data inputs. 776 bool no_dead_loop = true; 777 uint cnt = n->req(); 778 for (uint i = 1; i < cnt && no_dead_loop; i++) { 779 Node *in = n->in(i); 780 if (in == n) { 781 no_dead_loop = false; 782 } else if (in != nullptr && !in->is_dead_loop_safe()) { 783 uint icnt = in->req(); 784 for (uint j = 1; j < icnt && no_dead_loop; j++) { 785 if (in->in(j) == n || in->in(j) == in) 786 no_dead_loop = false; 787 } 788 } 789 } 790 if (!no_dead_loop) n->dump_bfs(100,0,"#"); 791 assert(no_dead_loop, "dead loop detected"); 792 } 793 } 794 795 796 /** 797 * Dumps information that can help to debug the problem. A debug 798 * build fails with an assert. 799 */ 800 void PhaseGVN::dump_infinite_loop_info(Node* n, const char* where) { 801 n->dump(4); 802 assert(false, "infinite loop in %s", where); 803 } 804 #endif 805 806 //============================================================================= 807 //------------------------------PhaseIterGVN----------------------------------- 808 // Initialize with previous PhaseIterGVN info; used by PhaseCCP 809 PhaseIterGVN::PhaseIterGVN(PhaseIterGVN* igvn) : _delay_transform(igvn->_delay_transform), 810 _worklist(*C->igvn_worklist()) 811 { 812 _iterGVN = true; 813 assert(&_worklist == &igvn->_worklist, "sanity"); 814 } 815 816 //------------------------------PhaseIterGVN----------------------------------- 817 // Initialize with previous PhaseGVN info from Parser 818 PhaseIterGVN::PhaseIterGVN(PhaseGVN* gvn) : _delay_transform(false), 819 _worklist(*C->igvn_worklist()) 820 { 821 _iterGVN = true; 822 uint max; 823 824 // Dead nodes in the hash table inherited from GVN were not treated as 825 // roots during def-use info creation; hence they represent an invisible 826 // use. Clear them out. 827 max = _table.size(); 828 for( uint i = 0; i < max; ++i ) { 829 Node *n = _table.at(i); 830 if(n != nullptr && n != _table.sentinel() && n->outcnt() == 0) { 831 if( n->is_top() ) continue; 832 // If remove_useless_nodes() has run, we expect no such nodes left. 833 assert(false, "remove_useless_nodes missed this node"); 834 hash_delete(n); 835 } 836 } 837 838 // Any Phis or Regions on the worklist probably had uses that could not 839 // make more progress because the uses were made while the Phis and Regions 840 // were in half-built states. Put all uses of Phis and Regions on worklist. 841 max = _worklist.size(); 842 for( uint j = 0; j < max; j++ ) { 843 Node *n = _worklist.at(j); 844 uint uop = n->Opcode(); 845 if( uop == Op_Phi || uop == Op_Region || 846 n->is_Type() || 847 n->is_Mem() ) 848 add_users_to_worklist(n); 849 } 850 } 851 852 void PhaseIterGVN::shuffle_worklist() { 853 if (_worklist.size() < 2) return; 854 for (uint i = _worklist.size() - 1; i >= 1; i--) { 855 uint j = C->random() % (i + 1); 856 swap(_worklist.adr()[i], _worklist.adr()[j]); 857 } 858 } 859 860 #ifndef PRODUCT 861 void PhaseIterGVN::verify_step(Node* n) { 862 if (is_verify_def_use()) { 863 ResourceMark rm; 864 VectorSet visited; 865 Node_List worklist; 866 867 _verify_window[_verify_counter % _verify_window_size] = n; 868 ++_verify_counter; 869 if (C->unique() < 1000 || 0 == _verify_counter % (C->unique() < 10000 ? 10 : 100)) { 870 ++_verify_full_passes; 871 worklist.push(C->root()); 872 Node::verify(-1, visited, worklist); 873 return; 874 } 875 for (int i = 0; i < _verify_window_size; i++) { 876 Node* n = _verify_window[i]; 877 if (n == nullptr) { 878 continue; 879 } 880 if (n->in(0) == NodeSentinel) { // xform_idom 881 _verify_window[i] = n->in(1); 882 --i; 883 continue; 884 } 885 // Typical fanout is 1-2, so this call visits about 6 nodes. 886 if (!visited.test_set(n->_idx)) { 887 worklist.push(n); 888 } 889 } 890 Node::verify(4, visited, worklist); 891 } 892 } 893 894 void PhaseIterGVN::trace_PhaseIterGVN(Node* n, Node* nn, const Type* oldtype) { 895 const Type* newtype = type_or_null(n); 896 if (nn != n || oldtype != newtype) { 897 C->print_method(PHASE_AFTER_ITER_GVN_STEP, 4, n); 898 } 899 if (TraceIterativeGVN) { 900 uint wlsize = _worklist.size(); 901 if (nn != n) { 902 // print old node 903 tty->print("< "); 904 if (oldtype != newtype && oldtype != nullptr) { 905 oldtype->dump(); 906 } 907 do { tty->print("\t"); } while (tty->position() < 16); 908 tty->print("<"); 909 n->dump(); 910 } 911 if (oldtype != newtype || nn != n) { 912 // print new node and/or new type 913 if (oldtype == nullptr) { 914 tty->print("* "); 915 } else if (nn != n) { 916 tty->print("> "); 917 } else { 918 tty->print("= "); 919 } 920 if (newtype == nullptr) { 921 tty->print("null"); 922 } else { 923 newtype->dump(); 924 } 925 do { tty->print("\t"); } while (tty->position() < 16); 926 nn->dump(); 927 } 928 if (Verbose && wlsize < _worklist.size()) { 929 tty->print(" Push {"); 930 while (wlsize != _worklist.size()) { 931 Node* pushed = _worklist.at(wlsize++); 932 tty->print(" %d", pushed->_idx); 933 } 934 tty->print_cr(" }"); 935 } 936 if (nn != n) { 937 // ignore n, it might be subsumed 938 verify_step((Node*) nullptr); 939 } 940 } 941 } 942 943 void PhaseIterGVN::init_verifyPhaseIterGVN() { 944 _verify_counter = 0; 945 _verify_full_passes = 0; 946 for (int i = 0; i < _verify_window_size; i++) { 947 _verify_window[i] = nullptr; 948 } 949 #ifdef ASSERT 950 // Verify that all modified nodes are on _worklist 951 Unique_Node_List* modified_list = C->modified_nodes(); 952 while (modified_list != nullptr && modified_list->size()) { 953 Node* n = modified_list->pop(); 954 if (!n->is_Con() && !_worklist.member(n)) { 955 n->dump(); 956 fatal("modified node is not on IGVN._worklist"); 957 } 958 } 959 #endif 960 } 961 962 void PhaseIterGVN::verify_PhaseIterGVN() { 963 #ifdef ASSERT 964 // Verify nodes with changed inputs. 965 Unique_Node_List* modified_list = C->modified_nodes(); 966 while (modified_list != nullptr && modified_list->size()) { 967 Node* n = modified_list->pop(); 968 if (!n->is_Con()) { // skip Con nodes 969 n->dump(); 970 fatal("modified node was not processed by IGVN.transform_old()"); 971 } 972 } 973 #endif 974 975 C->verify_graph_edges(); 976 if (is_verify_def_use() && PrintOpto) { 977 if (_verify_counter == _verify_full_passes) { 978 tty->print_cr("VerifyIterativeGVN: %d transforms and verify passes", 979 (int) _verify_full_passes); 980 } else { 981 tty->print_cr("VerifyIterativeGVN: %d transforms, %d full verify passes", 982 (int) _verify_counter, (int) _verify_full_passes); 983 } 984 } 985 986 #ifdef ASSERT 987 if (modified_list != nullptr) { 988 while (modified_list->size() > 0) { 989 Node* n = modified_list->pop(); 990 n->dump(); 991 assert(false, "VerifyIterativeGVN: new modified node was added"); 992 } 993 } 994 995 verify_optimize(); 996 #endif 997 } 998 #endif /* PRODUCT */ 999 1000 #ifdef ASSERT 1001 /** 1002 * Dumps information that can help to debug the problem. A debug 1003 * build fails with an assert. 1004 */ 1005 void PhaseIterGVN::dump_infinite_loop_info(Node* n, const char* where) { 1006 n->dump(4); 1007 _worklist.dump(); 1008 assert(false, "infinite loop in %s", where); 1009 } 1010 1011 /** 1012 * Prints out information about IGVN if the 'verbose' option is used. 1013 */ 1014 void PhaseIterGVN::trace_PhaseIterGVN_verbose(Node* n, int num_processed) { 1015 if (TraceIterativeGVN && Verbose) { 1016 tty->print(" Pop "); 1017 n->dump(); 1018 if ((num_processed % 100) == 0) { 1019 _worklist.print_set(); 1020 } 1021 } 1022 } 1023 #endif /* ASSERT */ 1024 1025 void PhaseIterGVN::optimize() { 1026 DEBUG_ONLY(uint num_processed = 0;) 1027 NOT_PRODUCT(init_verifyPhaseIterGVN();) 1028 C->print_method(PHASE_BEFORE_ITER_GVN, 3); 1029 if (StressIGVN) { 1030 shuffle_worklist(); 1031 } 1032 1033 uint loop_count = 0; 1034 // Pull from worklist and transform the node. If the node has changed, 1035 // update edge info and put uses on worklist. 1036 while(_worklist.size()) { 1037 if (C->check_node_count(NodeLimitFudgeFactor * 2, "Out of nodes")) { 1038 C->print_method(PHASE_AFTER_ITER_GVN, 3); 1039 return; 1040 } 1041 Node* n = _worklist.pop(); 1042 if (loop_count >= K * C->live_nodes()) { 1043 DEBUG_ONLY(dump_infinite_loop_info(n, "PhaseIterGVN::optimize");) 1044 C->record_method_not_compilable("infinite loop in PhaseIterGVN::optimize"); 1045 C->print_method(PHASE_AFTER_ITER_GVN, 3); 1046 return; 1047 } 1048 DEBUG_ONLY(trace_PhaseIterGVN_verbose(n, num_processed++);) 1049 if (n->outcnt() != 0) { 1050 NOT_PRODUCT(const Type* oldtype = type_or_null(n)); 1051 // Do the transformation 1052 Node* nn = transform_old(n); 1053 NOT_PRODUCT(trace_PhaseIterGVN(n, nn, oldtype);) 1054 } else if (!n->is_top()) { 1055 remove_dead_node(n); 1056 } 1057 loop_count++; 1058 } 1059 NOT_PRODUCT(verify_PhaseIterGVN();) 1060 C->print_method(PHASE_AFTER_ITER_GVN, 3); 1061 } 1062 1063 #ifdef ASSERT 1064 void PhaseIterGVN::verify_optimize() { 1065 if (is_verify_Value()) { 1066 ResourceMark rm; 1067 Unique_Node_List worklist; 1068 bool failure = false; 1069 // BFS all nodes, starting at root 1070 worklist.push(C->root()); 1071 for (uint j = 0; j < worklist.size(); ++j) { 1072 Node* n = worklist.at(j); 1073 failure |= verify_node_value(n); 1074 // traverse all inputs and outputs 1075 for (uint i = 0; i < n->req(); i++) { 1076 if (n->in(i) != nullptr) { 1077 worklist.push(n->in(i)); 1078 } 1079 } 1080 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1081 worklist.push(n->fast_out(i)); 1082 } 1083 } 1084 // If we get this assert, check why the reported nodes were not processed again in IGVN. 1085 // We should either make sure that these nodes are properly added back to the IGVN worklist 1086 // in PhaseIterGVN::add_users_to_worklist to update them again or add an exception 1087 // in the verification code above if that is not possible for some reason (like Load nodes). 1088 assert(!failure, "Missed optimization opportunity in PhaseIterGVN"); 1089 } 1090 } 1091 1092 // Check that type(n) == n->Value(), return true if we have a failure. 1093 // We have a list of exceptions, see detailed comments in code. 1094 // (1) Integer "widen" changes, but the range is the same. 1095 // (2) LoadNode performs deep traversals. Load is not notified for changes far away. 1096 // (3) CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away. 1097 bool PhaseIterGVN::verify_node_value(Node* n) { 1098 // If we assert inside type(n), because the type is still a null, then maybe 1099 // the node never went through gvn.transform, which would be a bug. 1100 const Type* told = type(n); 1101 const Type* tnew = n->Value(this); 1102 if (told == tnew) { 1103 return false; 1104 } 1105 // Exception (1) 1106 // Integer "widen" changes, but range is the same. 1107 if (told->isa_integer(tnew->basic_type()) != nullptr) { // both either int or long 1108 const TypeInteger* t0 = told->is_integer(tnew->basic_type()); 1109 const TypeInteger* t1 = tnew->is_integer(tnew->basic_type()); 1110 if (t0->lo_as_long() == t1->lo_as_long() && 1111 t0->hi_as_long() == t1->hi_as_long()) { 1112 return false; // ignore integer widen 1113 } 1114 } 1115 // Exception (2) 1116 // LoadNode performs deep traversals. Load is not notified for changes far away. 1117 if (n->is_Load() && !told->singleton()) { 1118 // MemNode::can_see_stored_value looks up through many memory nodes, 1119 // which means we would need to notify modifications from far up in 1120 // the inputs all the way down to the LoadNode. We don't do that. 1121 return false; 1122 } 1123 // Exception (3) 1124 // CmpPNode performs deep traversals if it compares oopptr. CmpP is not notified for changes far away. 1125 if (n->Opcode() == Op_CmpP && type(n->in(1))->isa_oopptr() && type(n->in(2))->isa_oopptr()) { 1126 // SubNode::Value 1127 // CmpPNode::sub 1128 // MemNode::detect_ptr_independence 1129 // MemNode::all_controls_dominate 1130 // We find all controls of a pointer load, and see if they dominate the control of 1131 // an allocation. If they all dominate, we know the allocation is after (independent) 1132 // of the pointer load, and we can say the pointers are different. For this we call 1133 // n->dominates(sub, nlist) to check if controls n of the pointer load dominate the 1134 // control sub of the allocation. The problems is that sometimes dominates answers 1135 // false conservatively, and later it can determine that it is indeed true. Loops with 1136 // Region heads can lead to giving up, whereas LoopNodes can be skipped easier, and 1137 // so the traversal becomes more powerful. This is difficult to remidy, we would have 1138 // to notify the CmpP of CFG updates. Luckily, we recompute CmpP::Value during CCP 1139 // after loop-opts, so that should take care of many of these cases. 1140 return false; 1141 } 1142 tty->cr(); 1143 tty->print_cr("Missed Value optimization:"); 1144 n->dump_bfs(1, 0, ""); 1145 tty->print_cr("Current type:"); 1146 told->dump_on(tty); 1147 tty->cr(); 1148 tty->print_cr("Optimized type:"); 1149 tnew->dump_on(tty); 1150 tty->cr(); 1151 return true; 1152 } 1153 #endif 1154 1155 /** 1156 * Register a new node with the optimizer. Update the types array, the def-use 1157 * info. Put on worklist. 1158 */ 1159 Node* PhaseIterGVN::register_new_node_with_optimizer(Node* n, Node* orig) { 1160 set_type_bottom(n); 1161 _worklist.push(n); 1162 if (orig != nullptr) C->copy_node_notes_to(n, orig); 1163 return n; 1164 } 1165 1166 //------------------------------transform-------------------------------------- 1167 // Non-recursive: idealize Node 'n' with respect to its inputs and its value 1168 Node *PhaseIterGVN::transform( Node *n ) { 1169 // If brand new node, make space in type array, and give it a type. 1170 ensure_type_or_null(n); 1171 if (type_or_null(n) == nullptr) { 1172 set_type_bottom(n); 1173 } 1174 1175 if (_delay_transform) { 1176 // Add the node to the worklist but don't optimize for now 1177 _worklist.push(n); 1178 return n; 1179 } 1180 1181 return transform_old(n); 1182 } 1183 1184 Node *PhaseIterGVN::transform_old(Node* n) { 1185 NOT_PRODUCT(set_transforms()); 1186 // Remove 'n' from hash table in case it gets modified 1187 _table.hash_delete(n); 1188 #ifdef ASSERT 1189 if (is_verify_def_use()) { 1190 assert(!_table.find_index(n->_idx), "found duplicate entry in table"); 1191 } 1192 #endif 1193 1194 // Allow Bool -> Cmp idealisation in late inlining intrinsics that return a bool 1195 if (n->is_Cmp()) { 1196 add_users_to_worklist(n); 1197 } 1198 1199 // Apply the Ideal call in a loop until it no longer applies 1200 Node* k = n; 1201 DEBUG_ONLY(dead_loop_check(k);) 1202 DEBUG_ONLY(bool is_new = (k->outcnt() == 0);) 1203 C->remove_modified_node(k); 1204 Node* i = apply_ideal(k, /*can_reshape=*/true); 1205 assert(i != k || is_new || i->outcnt() > 0, "don't return dead nodes"); 1206 #ifndef PRODUCT 1207 verify_step(k); 1208 #endif 1209 1210 DEBUG_ONLY(uint loop_count = 1;) 1211 while (i != nullptr) { 1212 #ifdef ASSERT 1213 if (loop_count >= K + C->live_nodes()) { 1214 dump_infinite_loop_info(i, "PhaseIterGVN::transform_old"); 1215 } 1216 #endif 1217 assert((i->_idx >= k->_idx) || i->is_top(), "Idealize should return new nodes, use Identity to return old nodes"); 1218 // Made a change; put users of original Node on worklist 1219 add_users_to_worklist(k); 1220 // Replacing root of transform tree? 1221 if (k != i) { 1222 // Make users of old Node now use new. 1223 subsume_node(k, i); 1224 k = i; 1225 } 1226 DEBUG_ONLY(dead_loop_check(k);) 1227 // Try idealizing again 1228 DEBUG_ONLY(is_new = (k->outcnt() == 0);) 1229 C->remove_modified_node(k); 1230 i = apply_ideal(k, /*can_reshape=*/true); 1231 assert(i != k || is_new || (i->outcnt() > 0), "don't return dead nodes"); 1232 #ifndef PRODUCT 1233 verify_step(k); 1234 #endif 1235 DEBUG_ONLY(loop_count++;) 1236 } 1237 1238 // If brand new node, make space in type array. 1239 ensure_type_or_null(k); 1240 1241 // See what kind of values 'k' takes on at runtime 1242 const Type* t = k->Value(this); 1243 assert(t != nullptr, "value sanity"); 1244 1245 // Since I just called 'Value' to compute the set of run-time values 1246 // for this Node, and 'Value' is non-local (and therefore expensive) I'll 1247 // cache Value. Later requests for the local phase->type of this Node can 1248 // use the cached Value instead of suffering with 'bottom_type'. 1249 if (type_or_null(k) != t) { 1250 #ifndef PRODUCT 1251 inc_new_values(); 1252 set_progress(); 1253 #endif 1254 set_type(k, t); 1255 // If k is a TypeNode, capture any more-precise type permanently into Node 1256 k->raise_bottom_type(t); 1257 // Move users of node to worklist 1258 add_users_to_worklist(k); 1259 } 1260 // If 'k' computes a constant, replace it with a constant 1261 if (t->singleton() && !k->is_Con()) { 1262 NOT_PRODUCT(set_progress();) 1263 Node* con = makecon(t); // Make a constant 1264 add_users_to_worklist(k); 1265 subsume_node(k, con); // Everybody using k now uses con 1266 return con; 1267 } 1268 1269 // Now check for Identities 1270 i = k->Identity(this); // Look for a nearby replacement 1271 if (i != k) { // Found? Return replacement! 1272 NOT_PRODUCT(set_progress();) 1273 add_users_to_worklist(k); 1274 subsume_node(k, i); // Everybody using k now uses i 1275 return i; 1276 } 1277 1278 // Global Value Numbering 1279 i = hash_find_insert(k); // Check for pre-existing node 1280 if (i && (i != k)) { 1281 // Return the pre-existing node if it isn't dead 1282 NOT_PRODUCT(set_progress();) 1283 add_users_to_worklist(k); 1284 subsume_node(k, i); // Everybody using k now uses i 1285 return i; 1286 } 1287 1288 // Return Idealized original 1289 return k; 1290 } 1291 1292 //---------------------------------saturate------------------------------------ 1293 const Type* PhaseIterGVN::saturate(const Type* new_type, const Type* old_type, 1294 const Type* limit_type) const { 1295 return new_type->narrow(old_type); 1296 } 1297 1298 //------------------------------remove_globally_dead_node---------------------- 1299 // Kill a globally dead Node. All uses are also globally dead and are 1300 // aggressively trimmed. 1301 void PhaseIterGVN::remove_globally_dead_node( Node *dead ) { 1302 enum DeleteProgress { 1303 PROCESS_INPUTS, 1304 PROCESS_OUTPUTS 1305 }; 1306 ResourceMark rm; 1307 Node_Stack stack(32); 1308 stack.push(dead, PROCESS_INPUTS); 1309 1310 while (stack.is_nonempty()) { 1311 dead = stack.node(); 1312 if (dead->Opcode() == Op_SafePoint) { 1313 dead->as_SafePoint()->disconnect_from_root(this); 1314 } 1315 uint progress_state = stack.index(); 1316 assert(dead != C->root(), "killing root, eh?"); 1317 assert(!dead->is_top(), "add check for top when pushing"); 1318 NOT_PRODUCT( set_progress(); ) 1319 if (progress_state == PROCESS_INPUTS) { 1320 // After following inputs, continue to outputs 1321 stack.set_index(PROCESS_OUTPUTS); 1322 if (!dead->is_Con()) { // Don't kill cons but uses 1323 bool recurse = false; 1324 // Remove from hash table 1325 _table.hash_delete( dead ); 1326 // Smash all inputs to 'dead', isolating him completely 1327 for (uint i = 0; i < dead->req(); i++) { 1328 Node *in = dead->in(i); 1329 if (in != nullptr && in != C->top()) { // Points to something? 1330 int nrep = dead->replace_edge(in, nullptr, this); // Kill edges 1331 assert((nrep > 0), "sanity"); 1332 if (in->outcnt() == 0) { // Made input go dead? 1333 stack.push(in, PROCESS_INPUTS); // Recursively remove 1334 recurse = true; 1335 } else if (in->outcnt() == 1 && 1336 in->has_special_unique_user()) { 1337 _worklist.push(in->unique_out()); 1338 } else if (in->outcnt() <= 2 && dead->is_Phi()) { 1339 if (in->Opcode() == Op_Region) { 1340 _worklist.push(in); 1341 } else if (in->is_Store()) { 1342 DUIterator_Fast imax, i = in->fast_outs(imax); 1343 _worklist.push(in->fast_out(i)); 1344 i++; 1345 if (in->outcnt() == 2) { 1346 _worklist.push(in->fast_out(i)); 1347 i++; 1348 } 1349 assert(!(i < imax), "sanity"); 1350 } 1351 } else { 1352 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(this, in); 1353 } 1354 if (ReduceFieldZeroing && dead->is_Load() && i == MemNode::Memory && 1355 in->is_Proj() && in->in(0) != nullptr && in->in(0)->is_Initialize()) { 1356 // A Load that directly follows an InitializeNode is 1357 // going away. The Stores that follow are candidates 1358 // again to be captured by the InitializeNode. 1359 for (DUIterator_Fast jmax, j = in->fast_outs(jmax); j < jmax; j++) { 1360 Node *n = in->fast_out(j); 1361 if (n->is_Store()) { 1362 _worklist.push(n); 1363 } 1364 } 1365 } 1366 } // if (in != nullptr && in != C->top()) 1367 } // for (uint i = 0; i < dead->req(); i++) 1368 if (recurse) { 1369 continue; 1370 } 1371 } // if (!dead->is_Con()) 1372 } // if (progress_state == PROCESS_INPUTS) 1373 1374 // Aggressively kill globally dead uses 1375 // (Rather than pushing all the outs at once, we push one at a time, 1376 // plus the parent to resume later, because of the indefinite number 1377 // of edge deletions per loop trip.) 1378 if (dead->outcnt() > 0) { 1379 // Recursively remove output edges 1380 stack.push(dead->raw_out(0), PROCESS_INPUTS); 1381 } else { 1382 // Finished disconnecting all input and output edges. 1383 stack.pop(); 1384 // Remove dead node from iterative worklist 1385 _worklist.remove(dead); 1386 C->remove_useless_node(dead); 1387 } 1388 } // while (stack.is_nonempty()) 1389 } 1390 1391 //------------------------------subsume_node----------------------------------- 1392 // Remove users from node 'old' and add them to node 'nn'. 1393 void PhaseIterGVN::subsume_node( Node *old, Node *nn ) { 1394 if (old->Opcode() == Op_SafePoint) { 1395 old->as_SafePoint()->disconnect_from_root(this); 1396 } 1397 assert( old != hash_find(old), "should already been removed" ); 1398 assert( old != C->top(), "cannot subsume top node"); 1399 // Copy debug or profile information to the new version: 1400 C->copy_node_notes_to(nn, old); 1401 // Move users of node 'old' to node 'nn' 1402 for (DUIterator_Last imin, i = old->last_outs(imin); i >= imin; ) { 1403 Node* use = old->last_out(i); // for each use... 1404 // use might need re-hashing (but it won't if it's a new node) 1405 rehash_node_delayed(use); 1406 // Update use-def info as well 1407 // We remove all occurrences of old within use->in, 1408 // so as to avoid rehashing any node more than once. 1409 // The hash table probe swamps any outer loop overhead. 1410 uint num_edges = 0; 1411 for (uint jmax = use->len(), j = 0; j < jmax; j++) { 1412 if (use->in(j) == old) { 1413 use->set_req(j, nn); 1414 ++num_edges; 1415 } 1416 } 1417 i -= num_edges; // we deleted 1 or more copies of this edge 1418 } 1419 1420 // Search for instance field data PhiNodes in the same region pointing to the old 1421 // memory PhiNode and update their instance memory ids to point to the new node. 1422 if (old->is_Phi() && old->as_Phi()->type()->has_memory() && old->in(0) != nullptr) { 1423 Node* region = old->in(0); 1424 for (DUIterator_Fast imax, i = region->fast_outs(imax); i < imax; i++) { 1425 PhiNode* phi = region->fast_out(i)->isa_Phi(); 1426 if (phi != nullptr && phi->inst_mem_id() == (int)old->_idx) { 1427 phi->set_inst_mem_id((int)nn->_idx); 1428 } 1429 } 1430 } 1431 1432 // Smash all inputs to 'old', isolating him completely 1433 Node *temp = new Node(1); 1434 temp->init_req(0,nn); // Add a use to nn to prevent him from dying 1435 remove_dead_node( old ); 1436 temp->del_req(0); // Yank bogus edge 1437 if (nn != nullptr && nn->outcnt() == 0) { 1438 _worklist.push(nn); 1439 } 1440 #ifndef PRODUCT 1441 if (is_verify_def_use()) { 1442 for ( int i = 0; i < _verify_window_size; i++ ) { 1443 if ( _verify_window[i] == old ) 1444 _verify_window[i] = nn; 1445 } 1446 } 1447 #endif 1448 temp->destruct(this); // reuse the _idx of this little guy 1449 } 1450 1451 void PhaseIterGVN::replace_in_uses(Node* n, Node* m) { 1452 assert(n != nullptr, "sanity"); 1453 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1454 Node* u = n->fast_out(i); 1455 if (u != n) { 1456 rehash_node_delayed(u); 1457 int nb = u->replace_edge(n, m); 1458 --i, imax -= nb; 1459 } 1460 } 1461 assert(n->outcnt() == 0, "all uses must be deleted"); 1462 } 1463 1464 //------------------------------add_users_to_worklist-------------------------- 1465 void PhaseIterGVN::add_users_to_worklist0(Node* n, Unique_Node_List& worklist) { 1466 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1467 worklist.push(n->fast_out(i)); // Push on worklist 1468 } 1469 } 1470 1471 // Return counted loop Phi if as a counted loop exit condition, cmp 1472 // compares the induction variable with n 1473 static PhiNode* countedloop_phi_from_cmp(CmpNode* cmp, Node* n) { 1474 for (DUIterator_Fast imax, i = cmp->fast_outs(imax); i < imax; i++) { 1475 Node* bol = cmp->fast_out(i); 1476 for (DUIterator_Fast i2max, i2 = bol->fast_outs(i2max); i2 < i2max; i2++) { 1477 Node* iff = bol->fast_out(i2); 1478 if (iff->is_BaseCountedLoopEnd()) { 1479 BaseCountedLoopEndNode* cle = iff->as_BaseCountedLoopEnd(); 1480 if (cle->limit() == n) { 1481 PhiNode* phi = cle->phi(); 1482 if (phi != nullptr) { 1483 return phi; 1484 } 1485 } 1486 } 1487 } 1488 } 1489 return nullptr; 1490 } 1491 1492 void PhaseIterGVN::add_users_to_worklist(Node *n) { 1493 add_users_to_worklist0(n, _worklist); 1494 1495 Unique_Node_List& worklist = _worklist; 1496 // Move users of node to worklist 1497 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1498 Node* use = n->fast_out(i); // Get use 1499 add_users_of_use_to_worklist(n, use, worklist); 1500 } 1501 } 1502 1503 void PhaseIterGVN::add_users_of_use_to_worklist(Node* n, Node* use, Unique_Node_List& worklist) { 1504 if(use->is_Multi() || // Multi-definer? Push projs on worklist 1505 use->is_Store() ) // Enable store/load same address 1506 add_users_to_worklist0(use, worklist); 1507 1508 // If we changed the receiver type to a call, we need to revisit 1509 // the Catch following the call. It's looking for a non-null 1510 // receiver to know when to enable the regular fall-through path 1511 // in addition to the NullPtrException path. 1512 if (use->is_CallDynamicJava() && n == use->in(TypeFunc::Parms)) { 1513 Node* p = use->as_CallDynamicJava()->proj_out_or_null(TypeFunc::Control); 1514 if (p != nullptr) { 1515 add_users_to_worklist0(p, worklist); 1516 } 1517 } 1518 1519 uint use_op = use->Opcode(); 1520 if(use->is_Cmp()) { // Enable CMP/BOOL optimization 1521 add_users_to_worklist0(use, worklist); // Put Bool on worklist 1522 if (use->outcnt() > 0) { 1523 Node* bol = use->raw_out(0); 1524 if (bol->outcnt() > 0) { 1525 Node* iff = bol->raw_out(0); 1526 if (iff->outcnt() == 2) { 1527 // Look for the 'is_x2logic' pattern: "x ? : 0 : 1" and put the 1528 // phi merging either 0 or 1 onto the worklist 1529 Node* ifproj0 = iff->raw_out(0); 1530 Node* ifproj1 = iff->raw_out(1); 1531 if (ifproj0->outcnt() > 0 && ifproj1->outcnt() > 0) { 1532 Node* region0 = ifproj0->raw_out(0); 1533 Node* region1 = ifproj1->raw_out(0); 1534 if( region0 == region1 ) 1535 add_users_to_worklist0(region0, worklist); 1536 } 1537 } 1538 } 1539 } 1540 if (use_op == Op_CmpI || use_op == Op_CmpL) { 1541 Node* phi = countedloop_phi_from_cmp(use->as_Cmp(), n); 1542 if (phi != nullptr) { 1543 // Input to the cmp of a loop exit check has changed, thus 1544 // the loop limit may have changed, which can then change the 1545 // range values of the trip-count Phi. 1546 worklist.push(phi); 1547 } 1548 } 1549 if (use_op == Op_CmpI) { 1550 Node* cmp = use; 1551 Node* in1 = cmp->in(1); 1552 Node* in2 = cmp->in(2); 1553 // Notify CmpI / If pattern from CastIINode::Value (left pattern). 1554 // Must also notify if in1 is modified and possibly turns into X (right pattern). 1555 // 1556 // in1 in2 in1 in2 1557 // | | | | 1558 // +--- | --+ | | 1559 // | | | | | 1560 // CmpINode | CmpINode 1561 // | | | 1562 // BoolNode | BoolNode 1563 // | | OR | 1564 // IfNode | IfNode 1565 // | | | 1566 // IfProj | IfProj X 1567 // | | | | 1568 // CastIINode CastIINode 1569 // 1570 if (in1 != in2) { // if they are equal, the CmpI can fold them away 1571 if (in1 == n) { 1572 // in1 modified -> could turn into X -> do traversal based on right pattern. 1573 for (DUIterator_Fast i2max, i2 = cmp->fast_outs(i2max); i2 < i2max; i2++) { 1574 Node* bol = cmp->fast_out(i2); // For each Bool 1575 if (bol->is_Bool()) { 1576 for (DUIterator_Fast i3max, i3 = bol->fast_outs(i3max); i3 < i3max; i3++) { 1577 Node* iff = bol->fast_out(i3); // For each If 1578 if (iff->is_If()) { 1579 for (DUIterator_Fast i4max, i4 = iff->fast_outs(i4max); i4 < i4max; i4++) { 1580 Node* if_proj = iff->fast_out(i4); // For each IfProj 1581 assert(if_proj->is_IfProj(), "If only has IfTrue and IfFalse as outputs"); 1582 for (DUIterator_Fast i5max, i5 = if_proj->fast_outs(i5max); i5 < i5max; i5++) { 1583 Node* castii = if_proj->fast_out(i5); // For each CastII 1584 if (castii->is_CastII() && 1585 castii->as_CastII()->carry_dependency()) { 1586 worklist.push(castii); 1587 } 1588 } 1589 } 1590 } 1591 } 1592 } 1593 } 1594 } else { 1595 // Only in2 modified -> can assume X == in2 (left pattern). 1596 assert(n == in2, "only in2 modified"); 1597 // Find all CastII with input in1. 1598 for (DUIterator_Fast jmax, j = in1->fast_outs(jmax); j < jmax; j++) { 1599 Node* castii = in1->fast_out(j); 1600 if (castii->is_CastII() && castii->as_CastII()->carry_dependency()) { 1601 // Find If. 1602 if (castii->in(0) != nullptr && castii->in(0)->in(0) != nullptr && castii->in(0)->in(0)->is_If()) { 1603 Node* ifnode = castii->in(0)->in(0); 1604 // Check that if connects to the cmp 1605 if (ifnode->in(1) != nullptr && ifnode->in(1)->is_Bool() && ifnode->in(1)->in(1) == cmp) { 1606 worklist.push(castii); 1607 } 1608 } 1609 } 1610 } 1611 } 1612 } 1613 } 1614 } 1615 1616 // Inline type nodes can have other inline types as users. If an input gets 1617 // updated, make sure that inline type users get a chance for optimization. 1618 if (use->is_InlineType()) { 1619 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1620 Node* u = use->fast_out(i2); 1621 if (u->is_InlineType()) 1622 worklist.push(u); 1623 } 1624 } 1625 // If changed Cast input, notify down for Phi, Sub, and Xor - all do "uncast" 1626 // Patterns: 1627 // ConstraintCast+ -> Sub 1628 // ConstraintCast+ -> Phi 1629 // ConstraintCast+ -> Xor 1630 if (use->is_ConstraintCast()) { 1631 auto push_the_uses_to_worklist = [&](Node* n){ 1632 if (n->is_Phi() || n->is_Sub() || n->Opcode() == Op_XorI || n->Opcode() == Op_XorL) { 1633 worklist.push(n); 1634 } 1635 }; 1636 auto is_boundary = [](Node* n){ return !n->is_ConstraintCast(); }; 1637 use->visit_uses(push_the_uses_to_worklist, is_boundary); 1638 } 1639 // If changed LShift inputs, check RShift users for useless sign-ext 1640 if( use_op == Op_LShiftI ) { 1641 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1642 Node* u = use->fast_out(i2); 1643 if (u->Opcode() == Op_RShiftI) 1644 worklist.push(u); 1645 } 1646 } 1647 // If changed LShift inputs, check And users for shift and mask (And) operation 1648 if (use_op == Op_LShiftI || use_op == Op_LShiftL) { 1649 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1650 Node* u = use->fast_out(i2); 1651 if (u->Opcode() == Op_AndI || u->Opcode() == Op_AndL) { 1652 worklist.push(u); 1653 } 1654 } 1655 } 1656 // If changed AddI/SubI inputs, check CmpU for range check optimization. 1657 if (use_op == Op_AddI || use_op == Op_SubI) { 1658 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1659 Node* u = use->fast_out(i2); 1660 if (u->is_Cmp() && (u->Opcode() == Op_CmpU)) { 1661 worklist.push(u); 1662 } 1663 } 1664 } 1665 // If changed AddP inputs, check Stores for loop invariant 1666 if( use_op == Op_AddP ) { 1667 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1668 Node* u = use->fast_out(i2); 1669 if (u->is_Mem()) 1670 worklist.push(u); 1671 } 1672 } 1673 // If changed initialization activity, check dependent Stores 1674 if (use_op == Op_Allocate || use_op == Op_AllocateArray) { 1675 InitializeNode* init = use->as_Allocate()->initialization(); 1676 if (init != nullptr) { 1677 Node* imem = init->proj_out_or_null(TypeFunc::Memory); 1678 if (imem != nullptr) add_users_to_worklist0(imem, worklist); 1679 } 1680 } 1681 // If the ValidLengthTest input changes then the fallthrough path out of the AllocateArray may have become dead. 1682 // CatchNode::Value() is responsible for killing that path. The CatchNode has to be explicitly enqueued for igvn 1683 // to guarantee the change is not missed. 1684 if (use_op == Op_AllocateArray && n == use->in(AllocateNode::ValidLengthTest)) { 1685 Node* p = use->as_AllocateArray()->proj_out_or_null(TypeFunc::Control); 1686 if (p != nullptr) { 1687 add_users_to_worklist0(p, worklist); 1688 } 1689 } 1690 1691 if (use_op == Op_Initialize) { 1692 Node* imem = use->as_Initialize()->proj_out_or_null(TypeFunc::Memory); 1693 if (imem != nullptr) add_users_to_worklist0(imem, worklist); 1694 } 1695 // Loading the java mirror from a Klass requires two loads and the type 1696 // of the mirror load depends on the type of 'n'. See LoadNode::Value(). 1697 // LoadBarrier?(LoadP(LoadP(AddP(foo:Klass, #java_mirror)))) 1698 BarrierSetC2* bs = BarrierSet::barrier_set()->barrier_set_c2(); 1699 bool has_load_barrier_nodes = bs->has_load_barrier_nodes(); 1700 1701 if (use_op == Op_CastP2X) { 1702 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1703 Node* u = use->fast_out(i2); 1704 if (u->Opcode() == Op_AndX) { 1705 worklist.push(u); 1706 } 1707 } 1708 } 1709 if (use_op == Op_LoadP && use->bottom_type()->isa_rawptr()) { 1710 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1711 Node* u = use->fast_out(i2); 1712 const Type* ut = u->bottom_type(); 1713 if (u->Opcode() == Op_LoadP && ut->isa_instptr()) { 1714 if (has_load_barrier_nodes) { 1715 // Search for load barriers behind the load 1716 for (DUIterator_Fast i3max, i3 = u->fast_outs(i3max); i3 < i3max; i3++) { 1717 Node* b = u->fast_out(i3); 1718 if (bs->is_gc_barrier_node(b)) { 1719 worklist.push(b); 1720 } 1721 } 1722 } 1723 worklist.push(u); 1724 } 1725 } 1726 } 1727 // Give CallStaticJavaNode::remove_useless_allocation a chance to run 1728 if (use->is_Region()) { 1729 Node* c = use; 1730 do { 1731 c = c->unique_ctrl_out_or_null(); 1732 } while (c != nullptr && c->is_Region()); 1733 if (c != nullptr && c->is_CallStaticJava() && c->as_CallStaticJava()->uncommon_trap_request() != 0) { 1734 worklist.push(c); 1735 } 1736 } 1737 if (use->Opcode() == Op_OpaqueZeroTripGuard) { 1738 assert(use->outcnt() <= 1, "OpaqueZeroTripGuard can't be shared"); 1739 if (use->outcnt() == 1) { 1740 Node* cmp = use->unique_out(); 1741 worklist.push(cmp); 1742 } 1743 } 1744 } 1745 1746 /** 1747 * Remove the speculative part of all types that we know of 1748 */ 1749 void PhaseIterGVN::remove_speculative_types() { 1750 assert(UseTypeSpeculation, "speculation is off"); 1751 for (uint i = 0; i < _types.Size(); i++) { 1752 const Type* t = _types.fast_lookup(i); 1753 if (t != nullptr) { 1754 _types.map(i, t->remove_speculative()); 1755 } 1756 } 1757 _table.check_no_speculative_types(); 1758 } 1759 1760 // Check if the type of a divisor of a Div or Mod node includes zero. 1761 bool PhaseIterGVN::no_dependent_zero_check(Node* n) const { 1762 switch (n->Opcode()) { 1763 case Op_DivI: 1764 case Op_ModI: { 1765 // Type of divisor includes 0? 1766 if (type(n->in(2)) == Type::TOP) { 1767 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations. 1768 return false; 1769 } 1770 const TypeInt* type_divisor = type(n->in(2))->is_int(); 1771 return (type_divisor->_hi < 0 || type_divisor->_lo > 0); 1772 } 1773 case Op_DivL: 1774 case Op_ModL: { 1775 // Type of divisor includes 0? 1776 if (type(n->in(2)) == Type::TOP) { 1777 // 'n' is dead. Treat as if zero check is still there to avoid any further optimizations. 1778 return false; 1779 } 1780 const TypeLong* type_divisor = type(n->in(2))->is_long(); 1781 return (type_divisor->_hi < 0 || type_divisor->_lo > 0); 1782 } 1783 } 1784 return true; 1785 } 1786 1787 //============================================================================= 1788 #ifndef PRODUCT 1789 uint PhaseCCP::_total_invokes = 0; 1790 uint PhaseCCP::_total_constants = 0; 1791 #endif 1792 //------------------------------PhaseCCP--------------------------------------- 1793 // Conditional Constant Propagation, ala Wegman & Zadeck 1794 PhaseCCP::PhaseCCP( PhaseIterGVN *igvn ) : PhaseIterGVN(igvn) { 1795 NOT_PRODUCT( clear_constants(); ) 1796 assert( _worklist.size() == 0, "" ); 1797 analyze(); 1798 } 1799 1800 #ifndef PRODUCT 1801 //------------------------------~PhaseCCP-------------------------------------- 1802 PhaseCCP::~PhaseCCP() { 1803 inc_invokes(); 1804 _total_constants += count_constants(); 1805 } 1806 #endif 1807 1808 1809 #ifdef ASSERT 1810 void PhaseCCP::verify_type(Node* n, const Type* tnew, const Type* told) { 1811 if (tnew->meet(told) != tnew->remove_speculative()) { 1812 n->dump(1); 1813 tty->print("told = "); told->dump(); tty->cr(); 1814 tty->print("tnew = "); tnew->dump(); tty->cr(); 1815 fatal("Not monotonic"); 1816 } 1817 assert(!told->isa_int() || !tnew->isa_int() || told->is_int()->_widen <= tnew->is_int()->_widen, "widen increases"); 1818 assert(!told->isa_long() || !tnew->isa_long() || told->is_long()->_widen <= tnew->is_long()->_widen, "widen increases"); 1819 } 1820 #endif //ASSERT 1821 1822 // In this analysis, all types are initially set to TOP. We iteratively call Value() on all nodes of the graph until 1823 // we reach a fixed-point (i.e. no types change anymore). We start with a list that only contains the root node. Each time 1824 // a new type is set, we push all uses of that node back to the worklist (in some cases, we also push grandchildren 1825 // or nodes even further down back to the worklist because their type could change as a result of the current type 1826 // change). 1827 void PhaseCCP::analyze() { 1828 // Initialize all types to TOP, optimistic analysis 1829 for (uint i = 0; i < C->unique(); i++) { 1830 _types.map(i, Type::TOP); 1831 } 1832 1833 // CCP worklist is placed on a local arena, so that we can allow ResourceMarks on "Compile::current()->resource_arena()". 1834 // We also do not want to put the worklist on "Compile::current()->comp_arena()", as that one only gets de-allocated after 1835 // Compile is over. The local arena gets de-allocated at the end of its scope. 1836 ResourceArea local_arena(mtCompiler); 1837 Unique_Node_List worklist(&local_arena); 1838 DEBUG_ONLY(Unique_Node_List worklist_verify(&local_arena);) 1839 1840 // Push root onto worklist 1841 worklist.push(C->root()); 1842 1843 assert(_root_and_safepoints.size() == 0, "must be empty (unused)"); 1844 _root_and_safepoints.push(C->root()); 1845 1846 // Pull from worklist; compute new value; push changes out. 1847 // This loop is the meat of CCP. 1848 while (worklist.size() != 0) { 1849 Node* n = fetch_next_node(worklist); 1850 DEBUG_ONLY(worklist_verify.push(n);) 1851 if (n->is_SafePoint()) { 1852 // Make sure safepoints are processed by PhaseCCP::transform even if they are 1853 // not reachable from the bottom. Otherwise, infinite loops would be removed. 1854 _root_and_safepoints.push(n); 1855 } 1856 const Type* new_type = n->Value(this); 1857 if (new_type != type(n)) { 1858 DEBUG_ONLY(verify_type(n, new_type, type(n));) 1859 dump_type_and_node(n, new_type); 1860 set_type(n, new_type); 1861 push_child_nodes_to_worklist(worklist, n); 1862 } 1863 } 1864 DEBUG_ONLY(verify_analyze(worklist_verify);) 1865 } 1866 1867 #ifdef ASSERT 1868 // For every node n on verify list, check if type(n) == n->Value() 1869 // We have a list of exceptions, see comments in verify_node_value. 1870 void PhaseCCP::verify_analyze(Unique_Node_List& worklist_verify) { 1871 bool failure = false; 1872 while (worklist_verify.size()) { 1873 Node* n = worklist_verify.pop(); 1874 failure |= verify_node_value(n); 1875 } 1876 // If we get this assert, check why the reported nodes were not processed again in CCP. 1877 // We should either make sure that these nodes are properly added back to the CCP worklist 1878 // in PhaseCCP::push_child_nodes_to_worklist() to update their type or add an exception 1879 // in the verification code above if that is not possible for some reason (like Load nodes). 1880 assert(!failure, "PhaseCCP not at fixpoint: analysis result may be unsound."); 1881 } 1882 #endif 1883 1884 // Fetch next node from worklist to be examined in this iteration. 1885 Node* PhaseCCP::fetch_next_node(Unique_Node_List& worklist) { 1886 if (StressCCP) { 1887 return worklist.remove(C->random() % worklist.size()); 1888 } else { 1889 return worklist.pop(); 1890 } 1891 } 1892 1893 #ifndef PRODUCT 1894 void PhaseCCP::dump_type_and_node(const Node* n, const Type* t) { 1895 if (TracePhaseCCP) { 1896 t->dump(); 1897 do { 1898 tty->print("\t"); 1899 } while (tty->position() < 16); 1900 n->dump(); 1901 } 1902 } 1903 #endif 1904 1905 // We need to propagate the type change of 'n' to all its uses. Depending on the kind of node, additional nodes 1906 // (grandchildren or even further down) need to be revisited as their types could also be improved as a result 1907 // of the new type of 'n'. Push these nodes to the worklist. 1908 void PhaseCCP::push_child_nodes_to_worklist(Unique_Node_List& worklist, Node* n) const { 1909 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1910 Node* use = n->fast_out(i); 1911 push_if_not_bottom_type(worklist, use); 1912 push_more_uses(worklist, n, use); 1913 } 1914 } 1915 1916 void PhaseCCP::push_if_not_bottom_type(Unique_Node_List& worklist, Node* n) const { 1917 if (n->bottom_type() != type(n)) { 1918 worklist.push(n); 1919 } 1920 } 1921 1922 // For some nodes, we need to propagate the type change to grandchildren or even further down. 1923 // Add them back to the worklist. 1924 void PhaseCCP::push_more_uses(Unique_Node_List& worklist, Node* parent, const Node* use) const { 1925 push_phis(worklist, use); 1926 push_catch(worklist, use); 1927 push_cmpu(worklist, use); 1928 push_counted_loop_phi(worklist, parent, use); 1929 push_cast(worklist, use); 1930 push_loadp(worklist, use); 1931 push_and(worklist, parent, use); 1932 push_cast_ii(worklist, parent, use); 1933 push_opaque_zero_trip_guard(worklist, use); 1934 } 1935 1936 1937 // We must recheck Phis too if use is a Region. 1938 void PhaseCCP::push_phis(Unique_Node_List& worklist, const Node* use) const { 1939 if (use->is_Region()) { 1940 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1941 push_if_not_bottom_type(worklist, use->fast_out(i)); 1942 } 1943 } 1944 } 1945 1946 // If we changed the receiver type to a call, we need to revisit the Catch node following the call. It's looking for a 1947 // non-null receiver to know when to enable the regular fall-through path in addition to the NullPtrException path. 1948 // Same is true if the type of a ValidLengthTest input to an AllocateArrayNode changes. 1949 void PhaseCCP::push_catch(Unique_Node_List& worklist, const Node* use) { 1950 if (use->is_Call()) { 1951 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1952 Node* proj = use->fast_out(i); 1953 if (proj->is_Proj() && proj->as_Proj()->_con == TypeFunc::Control) { 1954 Node* catch_node = proj->find_out_with(Op_Catch); 1955 if (catch_node != nullptr) { 1956 worklist.push(catch_node); 1957 } 1958 } 1959 } 1960 } 1961 } 1962 1963 // CmpU nodes can get their type information from two nodes up in the graph (instead of from the nodes immediately 1964 // above). Make sure they are added to the worklist if nodes they depend on are updated since they could be missed 1965 // and get wrong types otherwise. 1966 void PhaseCCP::push_cmpu(Unique_Node_List& worklist, const Node* use) const { 1967 uint use_op = use->Opcode(); 1968 if (use_op == Op_AddI || use_op == Op_SubI) { 1969 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 1970 Node* cmpu = use->fast_out(i); 1971 if (cmpu->Opcode() == Op_CmpU) { 1972 // Got a CmpU which might need the new type information from node n. 1973 push_if_not_bottom_type(worklist, cmpu); 1974 } 1975 } 1976 } 1977 } 1978 1979 // If n is used in a counted loop exit condition, then the type of the counted loop's Phi depends on the type of 'n'. 1980 // Seem PhiNode::Value(). 1981 void PhaseCCP::push_counted_loop_phi(Unique_Node_List& worklist, Node* parent, const Node* use) { 1982 uint use_op = use->Opcode(); 1983 if (use_op == Op_CmpI || use_op == Op_CmpL) { 1984 PhiNode* phi = countedloop_phi_from_cmp(use->as_Cmp(), parent); 1985 if (phi != nullptr) { 1986 worklist.push(phi); 1987 } 1988 } 1989 } 1990 1991 void PhaseCCP::push_cast(Unique_Node_List& worklist, const Node* use) { 1992 uint use_op = use->Opcode(); 1993 if (use_op == Op_CastP2X) { 1994 for (DUIterator_Fast i2max, i2 = use->fast_outs(i2max); i2 < i2max; i2++) { 1995 Node* u = use->fast_out(i2); 1996 if (u->Opcode() == Op_AndX) { 1997 worklist.push(u); 1998 } 1999 } 2000 } 2001 } 2002 2003 // Loading the java mirror from a Klass requires two loads and the type of the mirror load depends on the type of 'n'. 2004 // See LoadNode::Value(). 2005 void PhaseCCP::push_loadp(Unique_Node_List& worklist, const Node* use) const { 2006 BarrierSetC2* barrier_set = BarrierSet::barrier_set()->barrier_set_c2(); 2007 bool has_load_barrier_nodes = barrier_set->has_load_barrier_nodes(); 2008 2009 if (use->Opcode() == Op_LoadP && use->bottom_type()->isa_rawptr()) { 2010 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 2011 Node* loadp = use->fast_out(i); 2012 const Type* ut = loadp->bottom_type(); 2013 if (loadp->Opcode() == Op_LoadP && ut->isa_instptr() && ut != type(loadp)) { 2014 if (has_load_barrier_nodes) { 2015 // Search for load barriers behind the load 2016 push_load_barrier(worklist, barrier_set, loadp); 2017 } 2018 worklist.push(loadp); 2019 } 2020 } 2021 } 2022 } 2023 2024 void PhaseCCP::push_load_barrier(Unique_Node_List& worklist, const BarrierSetC2* barrier_set, const Node* use) { 2025 for (DUIterator_Fast imax, i = use->fast_outs(imax); i < imax; i++) { 2026 Node* barrier_node = use->fast_out(i); 2027 if (barrier_set->is_gc_barrier_node(barrier_node)) { 2028 worklist.push(barrier_node); 2029 } 2030 } 2031 } 2032 2033 // AndI/L::Value() optimizes patterns similar to (v << 2) & 3 to zero if they are bitwise disjoint. 2034 // Add the AndI/L nodes back to the worklist to re-apply Value() in case the shift value changed. 2035 // Pattern: parent -> LShift (use) -> (ConstraintCast | ConvI2L)* -> And 2036 void PhaseCCP::push_and(Unique_Node_List& worklist, const Node* parent, const Node* use) const { 2037 uint use_op = use->Opcode(); 2038 if ((use_op == Op_LShiftI || use_op == Op_LShiftL) 2039 && use->in(2) == parent) { // is shift value (right-hand side of LShift) 2040 auto push_and_uses_to_worklist = [&](Node* n){ 2041 uint opc = n->Opcode(); 2042 if (opc == Op_AndI || opc == Op_AndL) { 2043 push_if_not_bottom_type(worklist, n); 2044 } 2045 }; 2046 auto is_boundary = [](Node* n) { 2047 return !(n->is_ConstraintCast() || n->Opcode() == Op_ConvI2L); 2048 }; 2049 use->visit_uses(push_and_uses_to_worklist, is_boundary); 2050 } 2051 } 2052 2053 // CastII::Value() optimizes CmpI/If patterns if the right input of the CmpI has a constant type. If the CastII input is 2054 // the same node as the left input into the CmpI node, the type of the CastII node can be improved accordingly. Add the 2055 // CastII node back to the worklist to re-apply Value() to either not miss this optimization or to undo it because it 2056 // cannot be applied anymore. We could have optimized the type of the CastII before but now the type of the right input 2057 // of the CmpI (i.e. 'parent') is no longer constant. The type of the CastII must be widened in this case. 2058 void PhaseCCP::push_cast_ii(Unique_Node_List& worklist, const Node* parent, const Node* use) const { 2059 if (use->Opcode() == Op_CmpI && use->in(2) == parent) { 2060 Node* other_cmp_input = use->in(1); 2061 for (DUIterator_Fast imax, i = other_cmp_input->fast_outs(imax); i < imax; i++) { 2062 Node* cast_ii = other_cmp_input->fast_out(i); 2063 if (cast_ii->is_CastII()) { 2064 push_if_not_bottom_type(worklist, cast_ii); 2065 } 2066 } 2067 } 2068 } 2069 2070 void PhaseCCP::push_opaque_zero_trip_guard(Unique_Node_List& worklist, const Node* use) const { 2071 if (use->Opcode() == Op_OpaqueZeroTripGuard) { 2072 push_if_not_bottom_type(worklist, use->unique_out()); 2073 } 2074 } 2075 2076 //------------------------------do_transform----------------------------------- 2077 // Top level driver for the recursive transformer 2078 void PhaseCCP::do_transform() { 2079 // Correct leaves of new-space Nodes; they point to old-space. 2080 C->set_root( transform(C->root())->as_Root() ); 2081 assert( C->top(), "missing TOP node" ); 2082 assert( C->root(), "missing root" ); 2083 } 2084 2085 //------------------------------transform-------------------------------------- 2086 // Given a Node in old-space, clone him into new-space. 2087 // Convert any of his old-space children into new-space children. 2088 Node *PhaseCCP::transform( Node *n ) { 2089 assert(n->is_Root(), "traversal must start at root"); 2090 assert(_root_and_safepoints.member(n), "root (n) must be in list"); 2091 2092 ResourceMark rm; 2093 // Map: old node idx -> node after CCP (or nullptr if not yet transformed or useless). 2094 Node_List node_map; 2095 // Pre-allocate to avoid frequent realloc 2096 GrowableArray <Node *> transform_stack(C->live_nodes() >> 1); 2097 // track all visited nodes, so that we can remove the complement 2098 Unique_Node_List useful; 2099 2100 // Initialize the traversal. 2101 // This CCP pass may prove that no exit test for a loop ever succeeds (i.e. the loop is infinite). In that case, 2102 // the logic below doesn't follow any path from Root to the loop body: there's at least one such path but it's proven 2103 // never taken (its type is TOP). As a consequence the node on the exit path that's input to Root (let's call it n) is 2104 // replaced by the top node and the inputs of that node n are not enqueued for further processing. If CCP only works 2105 // through the graph from Root, this causes the loop body to never be processed here even when it's not dead (that 2106 // is reachable from Root following its uses). To prevent that issue, transform() starts walking the graph from Root 2107 // and all safepoints. 2108 for (uint i = 0; i < _root_and_safepoints.size(); ++i) { 2109 Node* nn = _root_and_safepoints.at(i); 2110 Node* new_node = node_map[nn->_idx]; 2111 assert(new_node == nullptr, ""); 2112 new_node = transform_once(nn); // Check for constant 2113 node_map.map(nn->_idx, new_node); // Flag as having been cloned 2114 transform_stack.push(new_node); // Process children of cloned node 2115 useful.push(new_node); 2116 } 2117 2118 while (transform_stack.is_nonempty()) { 2119 Node* clone = transform_stack.pop(); 2120 uint cnt = clone->req(); 2121 for( uint i = 0; i < cnt; i++ ) { // For all inputs do 2122 Node *input = clone->in(i); 2123 if( input != nullptr ) { // Ignore nulls 2124 Node *new_input = node_map[input->_idx]; // Check for cloned input node 2125 if( new_input == nullptr ) { 2126 new_input = transform_once(input); // Check for constant 2127 node_map.map( input->_idx, new_input );// Flag as having been cloned 2128 transform_stack.push(new_input); // Process children of cloned node 2129 useful.push(new_input); 2130 } 2131 assert( new_input == clone->in(i), "insanity check"); 2132 } 2133 } 2134 } 2135 2136 // The above transformation might lead to subgraphs becoming unreachable from the 2137 // bottom while still being reachable from the top. As a result, nodes in that 2138 // subgraph are not transformed and their bottom types are not updated, leading to 2139 // an inconsistency between bottom_type() and type(). In rare cases, LoadNodes in 2140 // such a subgraph, might be re-enqueued for IGVN indefinitely by MemNode::Ideal_common 2141 // because their address type is inconsistent. Therefore, we aggressively remove 2142 // all useless nodes here even before PhaseIdealLoop::build_loop_late gets a chance 2143 // to remove them anyway. 2144 if (C->cached_top_node()) { 2145 useful.push(C->cached_top_node()); 2146 } 2147 C->update_dead_node_list(useful); 2148 remove_useless_nodes(useful.member_set()); 2149 _worklist.remove_useless_nodes(useful.member_set()); 2150 C->disconnect_useless_nodes(useful, _worklist); 2151 2152 Node* new_root = node_map[n->_idx]; 2153 assert(new_root->is_Root(), "transformed root node must be a root node"); 2154 return new_root; 2155 } 2156 2157 //------------------------------transform_once--------------------------------- 2158 // For PhaseCCP, transformation is IDENTITY unless Node computed a constant. 2159 Node *PhaseCCP::transform_once( Node *n ) { 2160 const Type *t = type(n); 2161 // Constant? Use constant Node instead 2162 if( t->singleton() ) { 2163 Node *nn = n; // Default is to return the original constant 2164 if( t == Type::TOP ) { 2165 // cache my top node on the Compile instance 2166 if( C->cached_top_node() == nullptr || C->cached_top_node()->in(0) == nullptr ) { 2167 C->set_cached_top_node(ConNode::make(Type::TOP)); 2168 set_type(C->top(), Type::TOP); 2169 } 2170 nn = C->top(); 2171 } 2172 if( !n->is_Con() ) { 2173 if( t != Type::TOP ) { 2174 nn = makecon(t); // ConNode::make(t); 2175 NOT_PRODUCT( inc_constants(); ) 2176 } else if( n->is_Region() ) { // Unreachable region 2177 // Note: nn == C->top() 2178 n->set_req(0, nullptr); // Cut selfreference 2179 bool progress = true; 2180 uint max = n->outcnt(); 2181 DUIterator i; 2182 while (progress) { 2183 progress = false; 2184 // Eagerly remove dead phis to avoid phis copies creation. 2185 for (i = n->outs(); n->has_out(i); i++) { 2186 Node* m = n->out(i); 2187 if (m->is_Phi()) { 2188 assert(type(m) == Type::TOP, "Unreachable region should not have live phis."); 2189 replace_node(m, nn); 2190 if (max != n->outcnt()) { 2191 progress = true; 2192 i = n->refresh_out_pos(i); 2193 max = n->outcnt(); 2194 } 2195 } 2196 } 2197 } 2198 } 2199 replace_node(n,nn); // Update DefUse edges for new constant 2200 } 2201 return nn; 2202 } 2203 2204 // If x is a TypeNode, capture any more-precise type permanently into Node 2205 if (t != n->bottom_type()) { 2206 hash_delete(n); // changing bottom type may force a rehash 2207 n->raise_bottom_type(t); 2208 _worklist.push(n); // n re-enters the hash table via the worklist 2209 } 2210 2211 // TEMPORARY fix to ensure that 2nd GVN pass eliminates null checks 2212 switch( n->Opcode() ) { 2213 case Op_CallStaticJava: // Give post-parse call devirtualization a chance 2214 case Op_CallDynamicJava: 2215 case Op_FastLock: // Revisit FastLocks for lock coarsening 2216 case Op_If: 2217 case Op_CountedLoopEnd: 2218 case Op_Region: 2219 case Op_Loop: 2220 case Op_CountedLoop: 2221 case Op_Conv2B: 2222 case Op_Opaque1: 2223 _worklist.push(n); 2224 break; 2225 default: 2226 break; 2227 } 2228 2229 return n; 2230 } 2231 2232 //---------------------------------saturate------------------------------------ 2233 const Type* PhaseCCP::saturate(const Type* new_type, const Type* old_type, 2234 const Type* limit_type) const { 2235 const Type* wide_type = new_type->widen(old_type, limit_type); 2236 if (wide_type != new_type) { // did we widen? 2237 // If so, we may have widened beyond the limit type. Clip it back down. 2238 new_type = wide_type->filter(limit_type); 2239 } 2240 return new_type; 2241 } 2242 2243 //------------------------------print_statistics------------------------------- 2244 #ifndef PRODUCT 2245 void PhaseCCP::print_statistics() { 2246 tty->print_cr("CCP: %d constants found: %d", _total_invokes, _total_constants); 2247 } 2248 #endif 2249 2250 2251 //============================================================================= 2252 #ifndef PRODUCT 2253 uint PhasePeephole::_total_peepholes = 0; 2254 #endif 2255 //------------------------------PhasePeephole---------------------------------- 2256 // Conditional Constant Propagation, ala Wegman & Zadeck 2257 PhasePeephole::PhasePeephole( PhaseRegAlloc *regalloc, PhaseCFG &cfg ) 2258 : PhaseTransform(Peephole), _regalloc(regalloc), _cfg(cfg) { 2259 NOT_PRODUCT( clear_peepholes(); ) 2260 } 2261 2262 #ifndef PRODUCT 2263 //------------------------------~PhasePeephole--------------------------------- 2264 PhasePeephole::~PhasePeephole() { 2265 _total_peepholes += count_peepholes(); 2266 } 2267 #endif 2268 2269 //------------------------------transform-------------------------------------- 2270 Node *PhasePeephole::transform( Node *n ) { 2271 ShouldNotCallThis(); 2272 return nullptr; 2273 } 2274 2275 //------------------------------do_transform----------------------------------- 2276 void PhasePeephole::do_transform() { 2277 bool method_name_not_printed = true; 2278 2279 // Examine each basic block 2280 for (uint block_number = 1; block_number < _cfg.number_of_blocks(); ++block_number) { 2281 Block* block = _cfg.get_block(block_number); 2282 bool block_not_printed = true; 2283 2284 for (bool progress = true; progress;) { 2285 progress = false; 2286 // block->end_idx() not valid after PhaseRegAlloc 2287 uint end_index = block->number_of_nodes(); 2288 for( uint instruction_index = end_index - 1; instruction_index > 0; --instruction_index ) { 2289 Node *n = block->get_node(instruction_index); 2290 if( n->is_Mach() ) { 2291 MachNode *m = n->as_Mach(); 2292 // check for peephole opportunities 2293 int result = m->peephole(block, instruction_index, &_cfg, _regalloc); 2294 if( result != -1 ) { 2295 #ifndef PRODUCT 2296 if( PrintOptoPeephole ) { 2297 // Print method, first time only 2298 if( C->method() && method_name_not_printed ) { 2299 C->method()->print_short_name(); tty->cr(); 2300 method_name_not_printed = false; 2301 } 2302 // Print this block 2303 if( Verbose && block_not_printed) { 2304 tty->print_cr("in block"); 2305 block->dump(); 2306 block_not_printed = false; 2307 } 2308 // Print the peephole number 2309 tty->print_cr("peephole number: %d", result); 2310 } 2311 inc_peepholes(); 2312 #endif 2313 // Set progress, start again 2314 progress = true; 2315 break; 2316 } 2317 } 2318 } 2319 } 2320 } 2321 } 2322 2323 //------------------------------print_statistics------------------------------- 2324 #ifndef PRODUCT 2325 void PhasePeephole::print_statistics() { 2326 tty->print_cr("Peephole: peephole rules applied: %d", _total_peepholes); 2327 } 2328 #endif 2329 2330 2331 //============================================================================= 2332 //------------------------------set_req_X-------------------------------------- 2333 void Node::set_req_X( uint i, Node *n, PhaseIterGVN *igvn ) { 2334 assert( is_not_dead(n), "can not use dead node"); 2335 assert( igvn->hash_find(this) != this, "Need to remove from hash before changing edges" ); 2336 Node *old = in(i); 2337 set_req(i, n); 2338 2339 // old goes dead? 2340 if( old ) { 2341 switch (old->outcnt()) { 2342 case 0: 2343 // Put into the worklist to kill later. We do not kill it now because the 2344 // recursive kill will delete the current node (this) if dead-loop exists 2345 if (!old->is_top()) 2346 igvn->_worklist.push( old ); 2347 break; 2348 case 1: 2349 if( old->is_Store() || old->has_special_unique_user() ) 2350 igvn->add_users_to_worklist( old ); 2351 break; 2352 case 2: 2353 if( old->is_Store() ) 2354 igvn->add_users_to_worklist( old ); 2355 if( old->Opcode() == Op_Region ) 2356 igvn->_worklist.push(old); 2357 break; 2358 case 3: 2359 if( old->Opcode() == Op_Region ) { 2360 igvn->_worklist.push(old); 2361 igvn->add_users_to_worklist( old ); 2362 } 2363 break; 2364 default: 2365 break; 2366 } 2367 2368 BarrierSet::barrier_set()->barrier_set_c2()->enqueue_useful_gc_barrier(igvn, old); 2369 } 2370 } 2371 2372 void Node::set_req_X(uint i, Node *n, PhaseGVN *gvn) { 2373 PhaseIterGVN* igvn = gvn->is_IterGVN(); 2374 if (igvn == nullptr) { 2375 set_req(i, n); 2376 return; 2377 } 2378 set_req_X(i, n, igvn); 2379 } 2380 2381 //-------------------------------replace_by----------------------------------- 2382 // Using def-use info, replace one node for another. Follow the def-use info 2383 // to all users of the OLD node. Then make all uses point to the NEW node. 2384 void Node::replace_by(Node *new_node) { 2385 assert(!is_top(), "top node has no DU info"); 2386 for (DUIterator_Last imin, i = last_outs(imin); i >= imin; ) { 2387 Node* use = last_out(i); 2388 uint uses_found = 0; 2389 for (uint j = 0; j < use->len(); j++) { 2390 if (use->in(j) == this) { 2391 if (j < use->req()) 2392 use->set_req(j, new_node); 2393 else use->set_prec(j, new_node); 2394 uses_found++; 2395 } 2396 } 2397 i -= uses_found; // we deleted 1 or more copies of this edge 2398 } 2399 } 2400 2401 //============================================================================= 2402 //----------------------------------------------------------------------------- 2403 void Type_Array::grow( uint i ) { 2404 if( !_max ) { 2405 _max = 1; 2406 _types = (const Type**)_a->Amalloc( _max * sizeof(Type*) ); 2407 _types[0] = nullptr; 2408 } 2409 uint old = _max; 2410 _max = next_power_of_2(i); 2411 _types = (const Type**)_a->Arealloc( _types, old*sizeof(Type*),_max*sizeof(Type*)); 2412 memset( &_types[old], 0, (_max-old)*sizeof(Type*) ); 2413 } 2414 2415 //------------------------------dump------------------------------------------- 2416 #ifndef PRODUCT 2417 void Type_Array::dump() const { 2418 uint max = Size(); 2419 for( uint i = 0; i < max; i++ ) { 2420 if( _types[i] != nullptr ) { 2421 tty->print(" %d\t== ", i); _types[i]->dump(); tty->cr(); 2422 } 2423 } 2424 } 2425 #endif